MLflow Tracing: instrument an LLM/GenAI application with autolog and view traces in the MLflow UI

Steps

1. Install MLflow with GenAI extras: pip install 'mlflow[genai]', then start a tracking server with mlflow server (or use mlflow ui) so traces have somewhere to log to.
2. In your application code, call the one-line autolog function for your framework before making calls, e.g. import mlflow; mlflow.openai.autolog() (swap 'openai' for 'langchain', 'litellm', 'anthropic', etc. per the supported integration).
3. Run your LLM/agent code as normal; MLflow automatically captures each call as a trace (inputs, outputs, latency, and intermediate steps) and logs it to the active MLflow Experiment without extra code.
4. For custom logic not covered by an autolog integration, wrap functions with the @mlflow.trace decorator or trace a block of code manually to capture it as a span inside the trace.
5. Open the MLflow UI (Tracking Server URL, Traces tab within an Experiment) to inspect the trace timeline, per-step inputs/outputs, and latency for each request.
6. For production workloads, install the lighter 'mlflow-tracing' SDK package instead of full mlflow, and enable asynchronous trace logging so tracing doesn't add latency to live requests.